// the find
mlabonne/llm-datasets
Curated list of datasets and tools for post-training.
A README-only awesome-list of instruction, preference, and reasoning datasets for LLM post-training (SFT/DPO), maintained by Maxime Labonne. Aimed at ML engineers assembling a fine-tuning data mix who don't want to start from a blank HuggingFace search.
Organized by actual use case (general, math, code, instruction-following, multilingual, agent/function-calling, real conversations, preference) instead of a flat link dump. Each row carries sample count, a thinking/reasoning-trace flag, and license notes where it matters, which saves a click into the dataset card. Entries are dated and current through April 2026, so it's tracking the newest large releases (Nemotron Post-Training v3, Dolci, SYNTHETIC-2) rather than going stale. Pairs the dataset tables with a tools section (scraping, filtering, generation, exploration), so it also doubles as a pipeline reference, not just a bibliography.
It's a static README, not a package or API — nothing to install, nothing to query, you're clicking through to HuggingFace for everything. License info is inconsistent: some rows flag CC-BY-NC or NVIDIA's custom license, most say nothing, so you still have to check before using anything commercially. No guidance on which overlapping general-purpose mixture to actually pick for a given model size or compute budget — several entries compete for the same slot with no tiebreaker. Quality signal is just the maintainer's one-line note; there's no dedup or overlap check across datasets, so blending a few of these blind risks duplicate samples.